By Bryan Knipe, founder of TriPinnacle Advisors
Ask a leadership team how much the company spends on AI and what it gets back, and the honest answer is usually a pause. Not because nobody is tracking it, but because everybody is tracking it differently. Each business unit has its own projects, its own vendors, its own way of writing a business case and its own idea of what "working" means. The money is real. The view of it is not.
I ran into this at its largest scale at UnitedHealth Group. As COO/CFO of a large analytics organization, I set up and ran the evaluation of AI and machine-learning investments across the enterprise. Decisions had been made business unit by business unit, with no enterprise-level reconciliation. There were hundreds of investments, and no way to put two of them side by side and say which one deserved the next dollar.
The same problem shows up in a fifty-person company with eight AI pilots and in a health system with eighty. The method for fixing it is the same. Here is how I approach it.
1. Build one inventory before you build one strategy
The first job is not a strategy deck. It is a list. Every AI and machine-learning initiative, in one place, described the same way. For each one I want to know:
- What decision or piece of work it changes, in plain words.
- Who owns it on the business side and who owns it on the technical side.
- What it costs today and what it will cost to run: people, vendors, compute and licenses.
- What stage it is at: idea, pilot, in production, or in production and nobody is sure it is used.
- What it was promised to deliver, and whether anyone has checked.
Expect this to take longer than it should. Some projects are buried inside vendor contracts. Some are a team's side work that never got a budget line. Some have been quietly abandoned but are still paying for compute. The inventory alone usually finds money: duplicate tools bought by different units, two teams solving the same problem, and pilots that ended but never got turned off.
2. Agree on what counts as an AI investment
This sounds like a definitional argument, and it is one worth having early. If one unit counts a vendor's AI feature as an AI investment and another does not, your totals mean nothing. I draw the line at spend where a model makes or shapes a decision or produces work product. That includes vendor tools with AI built in, internal models, generative AI applications and the data work that exists only to feed them.
Write the definition down. You will be asked about edge cases for as long as the process runs, and a written answer stops the same debate from happening twelve times.
3. Score every project the same way
Once everything is in one list, it needs one yardstick. The scoring method I developed came down to three questions, weighted to fit the company's priorities:
- Strategic alignment. Does this move something leadership has said matters this year, or is it interesting work looking for a reason?
- Cost. The full cost to build and to run, including the people who have to review its output. AI that saves analyst time but needs a reviewer on every result has not saved what the business case says.
- Value. What changes, in a unit the business already measures: hours, error rates, cycle time, revenue, avoided cost. Not "efficiency." Not "insights."
The scores matter less than the fact that every project is scored by the same people against the same questions. A consistent method that is roughly right beats a perfect method that each unit applies its own way. Projects that cannot answer the value question at all are not bad projects yet. They are unfinished business cases, and they go back to their owners.
4. Make the business case something you can check later
Most AI business cases fail in the same place: the benefit is described so loosely that no one can ever prove it did or did not happen. Before anything is funded, I ask for three numbers written down:
- The baseline: what the work costs or how it performs today, measured, not estimated.
- The target: what it will be after the AI is in place, and by when.
- The measurement: where the number will come from, and who will pull it.
If a team cannot name the baseline, the first phase of the project is measuring it. That is not a delay. It is the cheapest insurance you can buy against funding something for a year and still not knowing whether it worked.
5. Decide: fund, fix or stop
With one list and one score, the portfolio sorts itself into three groups.
Fund the projects that score well and have a business case you can check. Give them clear owners and the next milestone.
Fix the ones that might be valuable but are not set up to prove it: a missing baseline, an owner who left, two projects that should be one. Most of the portfolio usually lands here.
Stop the ones that score poorly or duplicate something better. This is the hardest decision politically and the one that frees the most money. It helps to stop projects on the evidence of the scoring, applied the same way to everyone, rather than one at a time as a judgment on a team.
At UnitedHealth Group, bringing the investments into one view is what made consolidation possible: AI investments were brought together enterprise-wide, and I built the process for setting an enterprise AI strategy and roadmap from that view. The strategy came after the inventory, not before it.
6. Measure value after the money is spent
Funding decisions get attention. Value realization rarely does. A project that was approved on a promised benefit should be measured against that benefit on a schedule, and the result should feed the next funding decision.
We built a custom application to track realized value against what each investment promised, along with the methodology behind it. You do not need custom software to start. A shared tracker with the baseline, target and actual for each project, reviewed quarterly by the same group that makes funding decisions, does most of the work. What matters is that the people who approve the money see whether it paid off.
7. Put governance next to the money, not after it
AI governance often runs as a separate review that projects pass through on their way to production. I prefer to run it alongside the investment decision, with Legal, Compliance and Information Security at the table from the start. In healthcare and life sciences especially, data rights, privacy and regulatory exposure can make a high-value project unworkable. It is far cheaper to learn that at the scoring stage than after the build. Before I left, I ran AI/ML governance for the portfolio in partnership with those teams, and served on the team that reviewed proposed federal AI regulations and gave UnitedHealth Group's feedback to the federal government. The lesson from both: governance that sits beside the funding decision speeds good projects up, because they arrive at production already cleared.
What this looks like in a smaller company
None of this requires an enterprise. In a mid-size company, the inventory might be fifteen rows in a spreadsheet and the scoring might be one meeting. The steps are the same: list everything, define it the same way, score it the same way, write down a baseline, and decide what to fund, fix or stop. The work takes weeks, not quarters, and it usually pays for itself in the tools and pilots it turns off.
If you have more AI projects than you can explain to your board, that is the place to start. It is also the kind of work I do: I come in part-time, build the inventory and the scoring with your team, and stay through the first round of decisions.